AI News: ChatGPT Ultrafast, Grok 4.6, 3 New Open-Source Models, and more!

0:00 This is one of those weeks in AI that felt like a year. So many things launched that I just have to go over all of them. We have ultra fast mode for chat GPT. We have anthropic embedding

0:11 watermarks in their output. And we have four separate model releases, three of which were open source. I'm going to go over all of it. The first story we're going over today is ultra fast mode in

0:22 chatbt. And it all starts with this tweet. This is by Paul Graham. Making your model faster is a very smart move when your customers are paying by the token. Very true. More tokens per

0:33 second, more spend. But it is also extremely valuable to have really fast agents. And so, let me show you the news. Just a few months ago, Cerebrris announced a partnership with OpenAI. And

0:44 if you're not familiar with Cerebrris, they are the fastest inference provider on the planet. They build their own custom chips. I interviewed their CEO. This is him right here holding their

0:54 giant wafer. And now we have GPT 5.6 Soul running at 14 times the speed. And yes, of course, it's going to cost extra, but this is incredible. Here's an example of how fast it really is because

1:10 it is stunning. Build a financial terminal style dashboard for financial analysts. On the left, we have GPT 5.6 Soul with Cerebrris. On the right, we have regular Soul. And here we go. This

1:22 is going to run at 15x speed. And just look how fast it's done. A minute 50 seconds for an entire financial terminal. And over here on the right side, it's still going. 7 minutes, 8

1:34 minutes, 9 minutes. And if you've used codeex, you know sometimes you put out a prompt and it lasts for 30, 60, even 90 minutes. So here's the difference. 12 minutes 20 seconds on the right side

1:46 versus 1 minute 50 seconds. This is really going to change how developers build things because right now I'm kicking off 10 agents in parallel because I know each of them is going to

1:58 take 30 minutes and the context switching to go back and forth between 10 different agents takes a toll and it's actually really hard to do that context switching. And now because it's

2:09 going to be running so much faster, maybe I'll only have two or three agents running at the same time and that's going to literally change how I work. Right now it's only in preview. I did

2:19 have a chance to try it out briefly and it is very very fast and it seems like the bottleneck is no longer the thinking of the model. It's actually the tool calling and other things that require

2:32 traditional code and CPUs. That's right. CPUs are back, baby. And now this tweet by Tibo, who's on the Codeex team from a couple weeks ago, makes a lot more sense. The next generation of models

2:45 need more than your laptop. So we have massive models running at crazy speeds and no longer are the GPUs or in this case the cerebrus chips the bottleneck. It is now your computer. Your computer

3:00 can't keep up with how fast these large language models are running. So very interesting. It seems like we're moving to cloud agents even more so than we already are. And next this is crazy.

3:11 Enthropic is starting to add watermarks to its model output. So you give it a prompt, it gives you text back and there's a hidden watermark that you can't see in it so it can identify

3:23 output generated by clawed models and I don't even know how they're doing it. Maybe they're adding special characters to the text. Maybe they're using certain words and certain patterns in the text.

3:35 Does that change the output? and anthropic details why they have to do this, which is crazy because I haven't heard of any other AI company having to do this, but apparently it's all about

3:46 this EU AI act, article 52, code of practice on transparency of AI generated content. And so they did this to comply with EU regulation. And so there's just so much wrong with this, including why

4:02 is the EU doing this? Why is Anthropic seemingly the only company doing this? And what are the downstream effects of Anthropic being able to track your output from their models? It just feels

4:17 wrong. So, after Matt recorded and as I was editing the video you're watching right now, Anthropic actually dropped a blog that adds some clarity to how Claude is going to be watermarking its

4:27 outputs. And the most important part about this blog is they go over what the actual watermark is going to look like. Large language models like Claude work by generating one word at a time. Each

4:36 time the model decides on the next word, it chooses among a list of potential candidates, ultimately selecting the most sensible or likely based on the preceding text. Watermarking uses low

4:46 stakes choices like these to leave a pattern in Claw's responses. That pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it. When watermarking

4:57 is used, choices are still made at random, but the source of the randomness is different. Instead of using an arbitrary random number generator to pick the next word, watermarking uses

5:06 the key and a few words that come before to settle what word the model should pick. Then they go on to clarify about when the watermark will actually be used. The watermark only applies to

5:15 words Claude chooses. When Claude proofreads text written by a person, what it gives back generally only has been lightly edited. Because nearly all the words are the persons, there's very

5:26 little, if anything, for the watermark to attach to. The more Claude writes, the more decisions it has to make, and the more space there is for a watermark. They also talk about watermarks. When

5:35 Claude writes code in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code, but by definition,

5:45 it will have negligible effects on the actual code produced. And lastly, they talk about user impact, which of course they say has no impact on the user. Obviously, they're going to say this,

5:56 but pretty much anytime you add anything to a model like guardrails or watermarks, it's going to affect the output in some way or another. And in fact, they actually admit this. They say

6:05 that the word choice that is watermarked is still made at random, but the source is different, which is literally going to affect your output, whether it's low stakes or not. So, will this actually

6:15 affect Claude's outputs in the long run? We'll just have to see because this takes effect with future Claude models. Now, back to Matt next. So much news came out from Cursor and Grock and

6:27 SpaceX AI and they're all one company now. Cursor is now officially formally part of SpaceX, which is really cool. This is one of the best acquisitions of all time and I'm going to show you why.

6:39 They just released Grock 4.6 and looking at where they were even just 6 months ago, this is a massive improvement. Now they are very competitive at the Frontier. It is not quite as good as GPT

6:52 5.6. six soul or Fable 5, but it is really close and they are quickly gaining. So, they have a brand new model. It is inexpensive. $2 per million input, $6 per million output. It is very

7:05 fast and it is powering their newest product, which I'll show you now. This is Grockbot, and it is such a fun product to use. This is a completely new way to interact with AI. It is basically

7:22 all the power of a coding agent but simplified greatly. So you don't see the code, you don't see thinking, you don't have to choose models or thinking efforts and it just seems to work. I've

7:36 been using it for about a week now and I can say definitively I'm moving a lot of my workflows over to Grockbot. It has a lot of really cool unique features like every thread is actually an individual

7:48 agent. It has a bunch of plugins that you can use like Slack and Google Docs and email and all these different things. Easily just plug into it. It does just work. Think of it like

7:58 OpenClaw but for a broader audience. A more simple version of OpenClaw, a more polished version of OpenClaw. It also allows agents to talk to each other, which is really cool. So, not only can a

8:09 single agent delegate and create other agents, but it actually stores all of their conversations between each other. So you could put multiple agents in a conversation and just let them work

8:19 towards some goal. Really interesting, very flexible, but very simple. I've really enjoyed it. Definitely try it out. Next, we get to the open-source models. This week, multiple major

8:31 open-source releases happened. Let me show you all of them. First, we have GLM 5.3. And yes, it is quite good. Once again, not at the absolute frontier. So you can

8:44 see on Terminal Bench, for example, a huge improvement over GLM 5.2 beating Kimmy K3, but not quite at the level of a Fable or a Soul. Same thing with Deep Suite, which is probably the most

8:56 accurate benchmark that I've seen to how people really use these different models in coding environments. 66.9 massive jump from JLM 5.2, but just behind Kimmy, Fable, and decently behind Soul.

9:10 So you can see across the board it is very very competitive very much like Grock 4.6 extremely good model very cheap not quite even with these other Frontier models and DeepSseek released

9:24 Deepseek V4 Pro once again here's Terminal Bench 87.9 as compared to Kimmy K388 and Fable 588 as well and we can see it's an incredibly capable model. I'm going to show you the prices. Very

9:41 inexpensive and just barely behind the top models like Soul and Fable. Most importantly, look at these prices during off peak hours, which is kind of cool. I've never really seen this off peak

9:54 peak. 66 per million input tokens for a cash miss. $1.98 per million output tokens. And then during peak hours, it's just about double that price. But look at this. When you hit the cache, it is 2

10:07 cents per million input tokens, which is just crazy. Think about how many tokens you really get, how much context you can pack into this model so inexpensively. But that's not it. I am so happy to

10:19 report that Meta is getting back in the open source game. Maybe they never left, but they definitely released a few models that were completely closed source. We now have Muse Glimmer, their

10:30 latest model, which is open weights open source. Now this is a smaller model meant to run on device. It is a 30 billion parameter model which is basically the size for fitting it on a

10:41 desktop GPU. So if you have like a 40 series or a 50 series that's probably where you're going to be able to fit a 30 billion parameter model. Here it is on terminal bench coming in at a 51 well

10:53 below those other models that I just showed you. But this isn't meant to compete at the absolute frontier. This is very much meant for ondevice AI. I'm just happy Meta's back in the game

11:03 shipping new models and especially shipping open models. And if you like keeping up with all of these types of news stories, go to forwardfuture.com, go subscribe. And last, a little feature

11:14 launch that I think went under the radar a little bit. OpenAI launched computer history, which is kind of a weird name, but fine. Basically what it is is it records the things you're doing on your

11:27 computer and it tries to figure out ways to automate it and before you throw your arms up and say, "Oh my god, that is a privacy nightmare." It is optin only. You can have fine grain controls over

11:40 what you share. So if you only want to share your spreadsheets, you can do that. If you only want to share your browser, you can do that. Or you just leave it off completely. And if this

11:48 feature sounds familiar, remember back when Microsoft launched recall. Recall watched everything you did, recorded it, and then let you easily remember things using AI. I have to admit, I was very

12:02 excited about that feature. But Microsoft got absolutely destroyed upon launch. Everybody threw their arms up and said, "This is just a security and a privacy nightmare. Who would want this?"

12:13 Honestly, I would. Like, I'm okay. It was all stored locally. They were doing the LLM search locally. So, I was totally fine with it. And I think I'll probably be okay with this. I'll

12:25 probably dip my toe in. Probably try sharing a couple apps with computer history. But the point is for Chat GPT to be more proactive in suggesting things that it can do for you,

12:36 automations that it can set up for you. And so, if you want to use it, please do. Seems like a cool feature, but of course I think a lot of people are going to be very cautious letting chatt record

12:46 all of their computer use. And Alex on my team I think summed it up best. Open AAI crazy new releases in the next few weeks. Gemini, we have Gemini 3.7 flash spotted four pro rumors meta open

12:58 sourcing everything anthropic. Here's a watermark for you. I think my favorite release this week was Grockbot. And if you want to learn more about it, I made an entire video. Go check that out right

13:08 here.